Over several years, every successive era of technology has incorporated greater computing power into smaller areas. The trend towards maximum performance in data center operations has been significantly accelerated by AI, particularly in AI factory deployments. Operators are striving to optimize every rack, resulting in a profound transformation.
Two years ago, racks that seemed excessive are now considered commonplace, with the upcoming increment already in progress. The narrative is evident in the figures. A widespread measurement is “modal rack density,” which represents the power consumption per rack that occurs most frequently.
By 2026, the figure increased to 1003 kW, compared to 9 kW in 2025, as per the 16th Annual Global Data Center Survey conducted by the Uptime Institute. The 11 kW profile represents a usual enterprise server room, which can be found in numerous data centers today, rather than an AI training cluster.
AI racks belong to a separate power category altogether. Nvidia’s GB300 NVL72, which is crucial for AI training clusters between 2025 and early 2026, demands a power output of up to 142 kW per rack, as per Nvidia’s NVL4003 AI Factory reference architecture.
